In this paper, we present a novel deep-learning framework that incorporates quantified uncertainty for predicting the mechanical properties of nanocomposite materials, specifically taking into account their morphology and composition. Due to the intricate microstructures of nanocomposites and their dynamic changes under diverse conditions, traditional methods, such as molecular dynamics simulations, often impose significant computational burdens. Our machine learning models, trained on comprehensive material datasets, provide a lower computational cost alternative, facilitating rapid exploration of design spaces and more reliable predictions. We employ both convolutional neural networks and feedforward neural networks for our predictions, training separate models for yield strength and ultimate tensile strength. Furthermore, we integrate uncertainty quantification into our models, thereby providing confidence intervals for our predictions and making them more reliable. This study paves the way for advancements in predicting the properties of nanocomposite materials and could potentially be expanded to cover a broad spectrum of materials in the future.
{"title":"Integrating uncertainty into deep learning models for enhanced prediction of nanocomposite materials’ mechanical properties","authors":"Yuheng Wang, Guang Lin, Shengfeng Yang","doi":"10.1063/5.0177062","DOIUrl":"https://doi.org/10.1063/5.0177062","url":null,"abstract":"In this paper, we present a novel deep-learning framework that incorporates quantified uncertainty for predicting the mechanical properties of nanocomposite materials, specifically taking into account their morphology and composition. Due to the intricate microstructures of nanocomposites and their dynamic changes under diverse conditions, traditional methods, such as molecular dynamics simulations, often impose significant computational burdens. Our machine learning models, trained on comprehensive material datasets, provide a lower computational cost alternative, facilitating rapid exploration of design spaces and more reliable predictions. We employ both convolutional neural networks and feedforward neural networks for our predictions, training separate models for yield strength and ultimate tensile strength. Furthermore, we integrate uncertainty quantification into our models, thereby providing confidence intervals for our predictions and making them more reliable. This study paves the way for advancements in predicting the properties of nanocomposite materials and could potentially be expanded to cover a broad spectrum of materials in the future.","PeriodicalId":502250,"journal":{"name":"APL Machine Learning","volume":"6 6","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-02-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"140441204","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Electrical-circuit schematics are a foundational tool in electrical engineering. A method that can automatically digitalize them is desirable since a knowledge base of such schematics could preserve their functional information as well as provide a database that one can mine to predict more operationally efficient electrical circuits using data analytics and machine learning. We present a workflow that contains a novel pattern-recognition methodology and a custom-trained Optical Character Recognition (OCR) model that can digitalize images of electrical-circuit schematics with minimal configuration. The pattern-recognition and OCR stages of the workflow yield 86.4% and 99.6% success rates, respectively. We also present an extendable option toward predictive circuit-design efficiencies, subject to a large database of images being available. Thereby, data gathered from our pattern-recognition workflow are used to draw network graphs, which are in turn employed to form matrix equations that contain the voltages and currents for all nodes in the circuit in terms of component values. These equations could be applied to a database of electrical-circuit schematics to predict new circuit designs or circuit modifications that offer greater operational efficiency. Alternatively, these network graphs could be converted into simulation programs with integrated circuit emphasis netlists to afford more accurate and computationally automated simulations.
{"title":"Digitizing images of electrical-circuit schematics","authors":"Charles R. Kelly, Jacqueline M. Cole","doi":"10.1063/5.0177755","DOIUrl":"https://doi.org/10.1063/5.0177755","url":null,"abstract":"Electrical-circuit schematics are a foundational tool in electrical engineering. A method that can automatically digitalize them is desirable since a knowledge base of such schematics could preserve their functional information as well as provide a database that one can mine to predict more operationally efficient electrical circuits using data analytics and machine learning. We present a workflow that contains a novel pattern-recognition methodology and a custom-trained Optical Character Recognition (OCR) model that can digitalize images of electrical-circuit schematics with minimal configuration. The pattern-recognition and OCR stages of the workflow yield 86.4% and 99.6% success rates, respectively. We also present an extendable option toward predictive circuit-design efficiencies, subject to a large database of images being available. Thereby, data gathered from our pattern-recognition workflow are used to draw network graphs, which are in turn employed to form matrix equations that contain the voltages and currents for all nodes in the circuit in terms of component values. These equations could be applied to a database of electrical-circuit schematics to predict new circuit designs or circuit modifications that offer greater operational efficiency. Alternatively, these network graphs could be converted into simulation programs with integrated circuit emphasis netlists to afford more accurate and computationally automated simulations.","PeriodicalId":502250,"journal":{"name":"APL Machine Learning","volume":"29 12","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-01-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"139595647","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Ning Lin, Jia Chen, Ruoyu Zhao, Yangu He, Kwunhang Wong, Qinru Qiu, Zhongrui Wang, J. J. Yang
Despite the significant progress made in deep learning on digital computers, their energy consumption and computational speed still fall short of meeting the standards for brain-like computing. To address these limitations, reservoir computing (RC) has been gaining increasing attention across communities of electronic devices, computing systems, and machine learning, notably with its in-memory or in-sensor implementation on the hardware–software co-design. Hardware regarded, in-memory or in-sensor computers leverage emerging electronic and optoelectronic devices for data processing right where the data are stored or sensed. This technology dramatically reduces the energy consumption from frequent data transfers between sensing, storage, and computational units. Software regarded, RC enables real-time edge learning thanks to its brain-inspired dynamic system with massive training complexity reduction. From this perspective, we survey recent advancements in in-memory/in-sensor RC, including algorithm designs, material and device development, and downstream applications in classification and regression problems, and discuss challenges and opportunities ahead in this emerging field.
{"title":"In-memory and in-sensor reservoir computing with memristive devices","authors":"Ning Lin, Jia Chen, Ruoyu Zhao, Yangu He, Kwunhang Wong, Qinru Qiu, Zhongrui Wang, J. J. Yang","doi":"10.1063/5.0174863","DOIUrl":"https://doi.org/10.1063/5.0174863","url":null,"abstract":"Despite the significant progress made in deep learning on digital computers, their energy consumption and computational speed still fall short of meeting the standards for brain-like computing. To address these limitations, reservoir computing (RC) has been gaining increasing attention across communities of electronic devices, computing systems, and machine learning, notably with its in-memory or in-sensor implementation on the hardware–software co-design. Hardware regarded, in-memory or in-sensor computers leverage emerging electronic and optoelectronic devices for data processing right where the data are stored or sensed. This technology dramatically reduces the energy consumption from frequent data transfers between sensing, storage, and computational units. Software regarded, RC enables real-time edge learning thanks to its brain-inspired dynamic system with massive training complexity reduction. From this perspective, we survey recent advancements in in-memory/in-sensor RC, including algorithm designs, material and device development, and downstream applications in classification and regression problems, and discuss challenges and opportunities ahead in this emerging field.","PeriodicalId":502250,"journal":{"name":"APL Machine Learning","volume":"79 19","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-01-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"139593489","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
János Gergő Fehérvári, Z. Balogh, Tímea Nóra Török, A. Halbritter
The commercial introduction of a novel electronic device is often preceded by a lengthy material optimization phase devoted to the suppression of device noise as much as possible. The emergence of novel computing architectures, however, triggers a paradigm shift in noise engineering, demonstrating that non-suppressed but properly tailored noise can be harvested as a computational resource in probabilistic computing schemes. Such a strategy was recently realized on the hardware level in memristive Hopfield neural networks, delivering fast and highly energy efficient optimization performance. Inspired by these achievements, we perform a thorough analysis of simulated memristive Hopfield neural networks relying on realistic noise characteristics acquired on various memristive devices. These characteristics highlight the possibility of orders of magnitude variations in the noise level depending on the material choice as well as on the resistance state (and the corresponding active region volume) of the devices. Our simulations separate the effects of various device non-idealities on the operation of the Hopfield neural network by investigating the role of the programming accuracy as well as the noise-type and noise amplitude of the ON and OFF states. Relying on these results, we propose optimized noise tailoring and noise annealing strategies, comparing the impact of internal noise to the effect of external perturbation injection schemes.
{"title":"Noise tailoring, noise annealing, and external perturbation injection strategies in memristive Hopfield neural networks","authors":"János Gergő Fehérvári, Z. Balogh, Tímea Nóra Török, A. Halbritter","doi":"10.1063/5.0173662","DOIUrl":"https://doi.org/10.1063/5.0173662","url":null,"abstract":"The commercial introduction of a novel electronic device is often preceded by a lengthy material optimization phase devoted to the suppression of device noise as much as possible. The emergence of novel computing architectures, however, triggers a paradigm shift in noise engineering, demonstrating that non-suppressed but properly tailored noise can be harvested as a computational resource in probabilistic computing schemes. Such a strategy was recently realized on the hardware level in memristive Hopfield neural networks, delivering fast and highly energy efficient optimization performance. Inspired by these achievements, we perform a thorough analysis of simulated memristive Hopfield neural networks relying on realistic noise characteristics acquired on various memristive devices. These characteristics highlight the possibility of orders of magnitude variations in the noise level depending on the material choice as well as on the resistance state (and the corresponding active region volume) of the devices. Our simulations separate the effects of various device non-idealities on the operation of the Hopfield neural network by investigating the role of the programming accuracy as well as the noise-type and noise amplitude of the ON and OFF states. Relying on these results, we propose optimized noise tailoring and noise annealing strategies, comparing the impact of internal noise to the effect of external perturbation injection schemes.","PeriodicalId":502250,"journal":{"name":"APL Machine Learning","volume":"1 4","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-01-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"139600546","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Zhuang Yi, Du Yin, Lang Wu, Gaoqiang Niu, Feige Wang
Metal–oxide–semiconductor (MOS) gas sensors are widely used for gas detection and monitoring. However, MOS gas sensors have always suffered from instability in the link between gas sensor data and the measured gas concentration. In this paper, we propose a novel deep learning approach that combines the surface state model and a Gated Recurrent Unit (GRU)-based regression to enhance the analysis of gas sensor data. The surface state model provides valuable insights into the microscopic surface processes underlying the conductivity response to pulse heating, while the GRU model effectively captures the temporal dependencies present in time-series data. The experimental results demonstrate that the theory guided model GRU+β outperforms the elementary GRU algorithm in terms of accuracy and astringent speed. The incorporation of the surface state model and the parameter rate enhances the model’s accuracy and provides valuable information for learning pulse-heated regression tasks with better generalization. This research exhibits superiority of integrating domain knowledge and deep learning techniques in the field of gas sensor data analysis. The proposed approach offers a practical framework for improving the understanding and prediction of gas concentrations, facilitating better decision-making in various practical applications.
{"title":"A deep learning approach for gas sensor data regression: Incorporating surface state model and GRU-based model","authors":"Zhuang Yi, Du Yin, Lang Wu, Gaoqiang Niu, Feige Wang","doi":"10.1063/5.0160983","DOIUrl":"https://doi.org/10.1063/5.0160983","url":null,"abstract":"Metal–oxide–semiconductor (MOS) gas sensors are widely used for gas detection and monitoring. However, MOS gas sensors have always suffered from instability in the link between gas sensor data and the measured gas concentration. In this paper, we propose a novel deep learning approach that combines the surface state model and a Gated Recurrent Unit (GRU)-based regression to enhance the analysis of gas sensor data. The surface state model provides valuable insights into the microscopic surface processes underlying the conductivity response to pulse heating, while the GRU model effectively captures the temporal dependencies present in time-series data. The experimental results demonstrate that the theory guided model GRU+β outperforms the elementary GRU algorithm in terms of accuracy and astringent speed. The incorporation of the surface state model and the parameter rate enhances the model’s accuracy and provides valuable information for learning pulse-heated regression tasks with better generalization. This research exhibits superiority of integrating domain knowledge and deep learning techniques in the field of gas sensor data analysis. The proposed approach offers a practical framework for improving the understanding and prediction of gas concentrations, facilitating better decision-making in various practical applications.","PeriodicalId":502250,"journal":{"name":"APL Machine Learning","volume":"29 8","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-01-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"139532348","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Kapil Dev Mahato, S. S. G. Kumar Das, Chandrashekhar Azad, U. Kumar
Fluorescent organic dyes are extensively used in the design and discovery of new materials, photovoltaic cells, light sensors, imaging applications, medicinal chemistry, drug design, energy harvesting technologies, dye and pigment industries, and pharmaceutical industries, among other things. However, designing and synthesizing new fluorescent organic dyes with desirable properties for specific applications requires knowledge of the chemical and physical properties of previously studied molecules. It is a difficult task for experimentalists to identify the photophysical properties of the required chemical molecule at negligible time and financial cost. For this purpose, machine learning-based models are a highly demanding technique for estimating photophysical properties and may be an alternative approach to density functional theory. In this study, we used 15 single models and proposed three different hybrid models to assess a dataset of 3066 organic materials for predicting photophysical properties. The performance of these models was evaluated using three evaluation parameters: mean absolute error, root mean squared error, and the coefficient of determination (R2) on the test-size data. All the proposed hybrid models achieved the highest accuracy (R2) of 97.28%, 95.19%, and 74.01% for predicting the absorption wavelengths, emission wavelengths, and quantum yields, respectively. These resultant outcomes of the proposed hybrid models are ∼1.9%, ∼2.7%, and ∼2.4% higher than the recently reported best models’ values in the same dataset for absorption wavelengths, emission wavelengths, and quantum yields, respectively. This research promotes the quick and accurate production of new fluorescent organic dyes with desirable photophysical properties for specific applications.
{"title":"Machine learning based hybrid ensemble models for prediction of organic dyes photophysical properties: Absorption wavelengths, emission wavelengths, and quantum yields","authors":"Kapil Dev Mahato, S. S. G. Kumar Das, Chandrashekhar Azad, U. Kumar","doi":"10.1063/5.0181294","DOIUrl":"https://doi.org/10.1063/5.0181294","url":null,"abstract":"Fluorescent organic dyes are extensively used in the design and discovery of new materials, photovoltaic cells, light sensors, imaging applications, medicinal chemistry, drug design, energy harvesting technologies, dye and pigment industries, and pharmaceutical industries, among other things. However, designing and synthesizing new fluorescent organic dyes with desirable properties for specific applications requires knowledge of the chemical and physical properties of previously studied molecules. It is a difficult task for experimentalists to identify the photophysical properties of the required chemical molecule at negligible time and financial cost. For this purpose, machine learning-based models are a highly demanding technique for estimating photophysical properties and may be an alternative approach to density functional theory. In this study, we used 15 single models and proposed three different hybrid models to assess a dataset of 3066 organic materials for predicting photophysical properties. The performance of these models was evaluated using three evaluation parameters: mean absolute error, root mean squared error, and the coefficient of determination (R2) on the test-size data. All the proposed hybrid models achieved the highest accuracy (R2) of 97.28%, 95.19%, and 74.01% for predicting the absorption wavelengths, emission wavelengths, and quantum yields, respectively. These resultant outcomes of the proposed hybrid models are ∼1.9%, ∼2.7%, and ∼2.4% higher than the recently reported best models’ values in the same dataset for absorption wavelengths, emission wavelengths, and quantum yields, respectively. This research promotes the quick and accurate production of new fluorescent organic dyes with desirable photophysical properties for specific applications.","PeriodicalId":502250,"journal":{"name":"APL Machine Learning","volume":"42 23","pages":""},"PeriodicalIF":0.0,"publicationDate":"2024-01-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"139382487","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Antonia Pavlidou, Xiangpeng Liang, Negin Ghahremani Arekhloo, Haobo Li, J. Marquetand, Hadi Heidari
Neuromuscular disorders (NMDs) affect various parts of a motor unit, such as the motor neuron, neuromuscular junction, and muscle fibers. Abnormal spontaneous activity (SA) is detected with electromyography (EMG) as an essential hallmark in diagnosing NMD, which causes fatigue, pain, and muscle weakness. Monitoring the effects of NMD calls for new smart devices to collect and classify EMG. Delay-based Reservoir Computing (DRC) is a neuromorphic algorithm with high efficiency in classifying sequential data. This work proposes a new DRC-based algorithm that provides a reference for medical education and training and a second opinion to clinicians to verify NMD diagnoses by detecting SA in muscles. With a sampling frequency of Fs = 64 kHz, we have classified SA with EMG signals of 1 s of muscle recordings. Furthermore, the DRC model of size N = 600 nodes has successfully detected SA signals against normal muscle activity with an accuracy of up to 90.7%. The potential of using neuromorphic processing approaches in point-of-care diagnostics, alongside the supervision of a clinician, provides a more comprehensive and reliable clinical profile. Our developed model benefits from the potential to be implemented in physical hardware to provide near-sensor edge computing.
{"title":"Spontaneous muscle activity classification with delay-based reservoir computing","authors":"Antonia Pavlidou, Xiangpeng Liang, Negin Ghahremani Arekhloo, Haobo Li, J. Marquetand, Hadi Heidari","doi":"10.1063/5.0160927","DOIUrl":"https://doi.org/10.1063/5.0160927","url":null,"abstract":"Neuromuscular disorders (NMDs) affect various parts of a motor unit, such as the motor neuron, neuromuscular junction, and muscle fibers. Abnormal spontaneous activity (SA) is detected with electromyography (EMG) as an essential hallmark in diagnosing NMD, which causes fatigue, pain, and muscle weakness. Monitoring the effects of NMD calls for new smart devices to collect and classify EMG. Delay-based Reservoir Computing (DRC) is a neuromorphic algorithm with high efficiency in classifying sequential data. This work proposes a new DRC-based algorithm that provides a reference for medical education and training and a second opinion to clinicians to verify NMD diagnoses by detecting SA in muscles. With a sampling frequency of Fs = 64 kHz, we have classified SA with EMG signals of 1 s of muscle recordings. Furthermore, the DRC model of size N = 600 nodes has successfully detected SA signals against normal muscle activity with an accuracy of up to 90.7%. The potential of using neuromorphic processing approaches in point-of-care diagnostics, alongside the supervision of a clinician, provides a more comprehensive and reliable clinical profile. Our developed model benefits from the potential to be implemented in physical hardware to provide near-sensor edge computing.","PeriodicalId":502250,"journal":{"name":"APL Machine Learning","volume":"18 1","pages":""},"PeriodicalIF":0.0,"publicationDate":"2023-11-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"139199974","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Recently, the incorporation of machine learning (ML) has heralded significant advancements in materials science. For instance, in spintronics, it has been shown that magnetic parameters, such as the Dzyaloshinskii–Moriya interaction, can be estimated from magnetic domain images using ML. Magnetic materials exhibit hysteresis, leading to numerous magnetic states with locally minimized energy (LME) even within a single sample. However, it remains uncertain whether these parameters can be derived from LME states. In our research, we explored the estimation of material parameters from an LME magnetic state using a convolutional neural network. We introduced a technique to manipulate LME magnetic states, combining the ac demagnetizing method with the magneto-optical Kerr effect. By applying this method, we generated multiple LME magnetic states from a single sample and successfully estimated its material composition. Our findings suggest that ML emphasizes not the global domain structures that are readily perceived by humans but the more subtle local domain structures that are often overlooked. Adopting this approach could potentially facilitate the estimation of magnetic parameters from any state observed in experiments, streamlining experimental processes in spintronics.
{"title":"Estimation of TbCo composition from local-minimum-energy magnetic images taken by magneto-optical Kerr effect microscope by using machine learning","authors":"Shiori Kuno, Shinji Deguchi, Satoshi Sumi, Hiroyuki Awano, Kenji Tanabe","doi":"10.1063/5.0160970","DOIUrl":"https://doi.org/10.1063/5.0160970","url":null,"abstract":"Recently, the incorporation of machine learning (ML) has heralded significant advancements in materials science. For instance, in spintronics, it has been shown that magnetic parameters, such as the Dzyaloshinskii–Moriya interaction, can be estimated from magnetic domain images using ML. Magnetic materials exhibit hysteresis, leading to numerous magnetic states with locally minimized energy (LME) even within a single sample. However, it remains uncertain whether these parameters can be derived from LME states. In our research, we explored the estimation of material parameters from an LME magnetic state using a convolutional neural network. We introduced a technique to manipulate LME magnetic states, combining the ac demagnetizing method with the magneto-optical Kerr effect. By applying this method, we generated multiple LME magnetic states from a single sample and successfully estimated its material composition. Our findings suggest that ML emphasizes not the global domain structures that are readily perceived by humans but the more subtle local domain structures that are often overlooked. Adopting this approach could potentially facilitate the estimation of magnetic parameters from any state observed in experiments, streamlining experimental processes in spintronics.","PeriodicalId":502250,"journal":{"name":"APL Machine Learning","volume":"4 1","pages":""},"PeriodicalIF":0.0,"publicationDate":"2023-11-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"139213306","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Haoyue Deng, Zhen Fan, Shuai Dong, Zhiwei Chen, Wenjie Li, Yihong Chen, Kun Liu, Ruiqiang Tao, G. Tian, Deyang Chen, M. Qin, Min Zeng, Xubing Lu, G. Zhou, Xingsen Gao, Junming Liu
Memristive neural networks have extensively been investigated for their capability in handling various artificial intelligence tasks. The training performance of memristive neural networks depends on the pulse scheme applied to the constituent memristors. However, the design of the pulse scheme in most previous studies was approached in an empirical manner or through a trial-and-error method. Here, we choose ferroelectric tunnel junction (FTJ) as a model memristor and demonstrate a physics-based predictive model for the pulse design to achieve high training performance. This predictive model comprises a physical model for FTJ that can adequately describe the polarization switching and memristive switching behaviors of the FTJ and an FTJ-based neural network that uses the long-term potentiation (LTP)/long-term depression (LTD) characteristics of the FTJ for the weight update. Simulation results based on the predictive model demonstrate that the LTP/LTD characteristics with a good trade-off between ON/OFF ratio, nonlinearity, and asymmetry can lead to high training accuracies for the FTJ-based neural network. Moreover, it is revealed that an amplitude-increasing pulse scheme may be the most favorable pulse scheme as it offers the widest ranges of pulse amplitudes and widths for achieving high accuracies. This study may provide useful guidance for the pulse design in the experimental development of high-performance memristive neural networks.
{"title":"A physics-based predictive model for pulse design to realize high-performance memristive neural networks","authors":"Haoyue Deng, Zhen Fan, Shuai Dong, Zhiwei Chen, Wenjie Li, Yihong Chen, Kun Liu, Ruiqiang Tao, G. Tian, Deyang Chen, M. Qin, Min Zeng, Xubing Lu, G. Zhou, Xingsen Gao, Junming Liu","doi":"10.1063/5.0180346","DOIUrl":"https://doi.org/10.1063/5.0180346","url":null,"abstract":"Memristive neural networks have extensively been investigated for their capability in handling various artificial intelligence tasks. The training performance of memristive neural networks depends on the pulse scheme applied to the constituent memristors. However, the design of the pulse scheme in most previous studies was approached in an empirical manner or through a trial-and-error method. Here, we choose ferroelectric tunnel junction (FTJ) as a model memristor and demonstrate a physics-based predictive model for the pulse design to achieve high training performance. This predictive model comprises a physical model for FTJ that can adequately describe the polarization switching and memristive switching behaviors of the FTJ and an FTJ-based neural network that uses the long-term potentiation (LTP)/long-term depression (LTD) characteristics of the FTJ for the weight update. Simulation results based on the predictive model demonstrate that the LTP/LTD characteristics with a good trade-off between ON/OFF ratio, nonlinearity, and asymmetry can lead to high training accuracies for the FTJ-based neural network. Moreover, it is revealed that an amplitude-increasing pulse scheme may be the most favorable pulse scheme as it offers the widest ranges of pulse amplitudes and widths for achieving high accuracies. This study may provide useful guidance for the pulse design in the experimental development of high-performance memristive neural networks.","PeriodicalId":502250,"journal":{"name":"APL Machine Learning","volume":"78 1","pages":""},"PeriodicalIF":0.0,"publicationDate":"2023-11-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"139222031","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：